The Navier-Stokes Singularity Rift: OpenAI's 88-Hour Claim and Two Mathematicians' Pushback

For ninety years no one fully tamed Navier-Stokes. Silicon Carne's twenty-minute episode uses the 19th-century fluid equations to explain why a Clay million-dollar question suddenly returned to headlines on 8 September with OpenAI's announcement; even the literary-minded host's bemusement shows why the dispute unsettled mathematicians and outsiders alike.

Why water and air still challenge mathematicians

The equations are Newton's F equals ma adapted to fluids. Navier and Stokes treat water or air not as individual molecules but as a continuum, writing how pressure, density and velocity evolve. The air over a wing, blood in a vessel or water around a hull all share the same core. In 1934 Jean Leray showed generalized solutions exist, yet whether they stay smooth remained open; in 2000 the Clay Institute listed it as one of seven Millennium problems with a $1 million prize, and only Poincaré, solved by Perelman who declined the award, has been closed so far.

Clay's question is conceptually sharp: can a perfectly smooth fluid at rest, even while viscosity tries to smooth everything, develop a singularity where speed grows without bound in finite time? No real fluid can move infinitely fast, so such a blowup would mean the model itself breaks down and one must track every particle individually. That is why the debate is not only abstract proof but also about the reliability of engineering's base model.

Mathematicians therefore tried a simpler ladder. The Euler equations that drop viscosity, the Boussinesq approximation and the porous-media equation were seen as steps toward Navier-Stokes. Diego Cordoba and Luis Martinez-Zoroa had run a program that first built blowups with rough forcing; the idea did not come from a language model. Buckmaster and Alpoge took that line and tried to move it from rough to smooth forcing.

A quiet year, then a month of acceleration

Tristan Buckmaster is a professor at NYU Courant, Levent Alpoge works as a mathematician at Anthropic; the pair collaborated for about a year without institutional agreement, as a personal project. They say they used Claude, Codex and especially GPT-5.6 Sol throughout, and Astra later for write-up and auditing. By their own account the first model-generated proof they exchanged was the worst they had ever read. After months maturing a smooth-forcing blowup for porous media, on 15 August they obtained smooth-forcing finite-time blowups for Boussinesq and for three-dimensional incompressible Euler, with Lean verification finished on 22 August; they themselves describe the Euler write-up as raw, machine-guided output and apologize to the community.

As they focused on turning the material into readable papers, a rumor accelerated everything. Because they had loaded drafts directly into Codex, the possibility arose that progress had leaked outside. Around 1 September OpenAI heard the chatter about this advance and pointed its internal model at all Clay problems. It probed Euler first, then scaled up and, according to its account, ran ten thousand agents in parallel and reached a result after 88 hours in total.

OpenAI's narrative goes like this: a model trained in late August and said to be significantly stronger than GPT-6 Astra first found an Euler blowup in about 50 hours with a thousand agents, then extended to Navier-Stokes in 11 hours with ten thousand agents, closing the four-day run in 88 hours. At a press briefing the company estimated a customer rerun would cost roughly $15 million; the claimed result is a smooth force applied to a fluid at rest, with finite energy throughout, that drives a finite-time singularity, backed by a Lean formalization, and presented as settling alternatives C and D in Clay's statement — not a blowup for the unforced viscous equation.

The overlapping calendar inflamed matters. Buckmaster says he learned on 3 September that information about their progress had reached OpenAI, and hours before OpenAI's announcement on 8 September he posted a detailed statement with email excerpts, avoiding a direct accusation but stressing he did not want a false narrative to settle. OpenAI replied the same day that it had not seen any work until it was public, that no user data was accessed, that de-identified usage might have helped improve models but could not be ruled out, and that the two proofs differ notably in method and precise result while calling the Buckmaster-Alpoge concurrent work remarkable.

No official verification yet, but Lean helps

The Clay Institute has not yet validated either proof and peer review has not started. Still, a Lean formalization turns mathematics into code that a machine checks step by step, sharply lowering the chance of hidden error. Outside experts such as Dallas Albritton of Wisconsin called knowing the answer to a guiding problem a huge threshold, while BBC, New Scientist and Science News noted both the substance and the controversial timing in parallel coverage.

This blowup did not arrive alone. In the last twelve months a model from OpenAI resolved a longstanding Erdos conjecture in May, Claude Fable 5 produced a counterexample to the near-century-old Jacobian conjecture, and Fermat's last theorem was formalized in just eleven days. Sébastien Bubeck and Venkat Chandrasekaran at OpenAI described the streak as a spectacular combination of the past year's arc, noting that pen-and-paper calculations needed for such problems have become mind-bogglingly intricate for humans.

What fuels the controversy is less correctness than credit and data stewardship. Keeping drafts in Codex was convenient but also left traces on the model provider's infrastructure; OpenAI's cautious line that de-identified usage may have helped improve models points exactly to that gray zone. As the Silicon Carne discussion noted, when ultra-rushed write-ups in a vibe-coding tempo sit next to a ten-thousand-agent mega-run, the question of how scientific labor is protected becomes as important as the technical detail.

The economic and geopolitical layer grew at the same pace. Pricing a single run at about $15 million suggests a career bet once spread over a decade can now shrink to a four-week computation bet; a lab that cannot afford the bill fears being left behind, recalling how researchers once migrated to universities that owned supercomputers. That is why voices in the discussion stressed that Europe must own proprietary models, not merely distribute others; as the hosts put it, the country that can allocate AI agents to researchers will attract the next generation of talent.

Silicon Carne's tone adds a human frame. The host who calls himself literary, the laughter breaking through a calm account of a million-dollar question, the analogy of a Sun Valley Netflix and jokes about small quarrels turning into big money all remind us this is a community story beyond an equation. The team returns in the final minutes to the same line: AI has clearly accelerated mathematics, whether ownership and the reputation economy stay human will be tested in the next proof.